From b8bea00855bf7fcece01c2108b896e4c164cc624 Mon Sep 17 00:00:00 2001
From: SunXiaoye <31361630+JingofXin@users.noreply.github.com>
Date: Tue, 3 Dec 2024 16:30:11 +0800
Subject: [PATCH] Fixed batch_size_per_device and batch_size misuse in LazyLLM
 (#377)

---
 lazyllm/components/finetune/llamafactory/sft.yaml | 2 +-
 lazyllm/tools/train_service/client.py             | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/lazyllm/components/finetune/llamafactory/sft.yaml b/lazyllm/components/finetune/llamafactory/sft.yaml
index 064a01c5..a615a426 100644
--- a/lazyllm/components/finetune/llamafactory/sft.yaml
+++ b/lazyllm/components/finetune/llamafactory/sft.yaml
@@ -78,7 +78,7 @@ per_device_train_batch_size: 1
 per_device_eval_batch_size: 1
 per_gpu_train_batch_size: null
 per_gpu_eval_batch_size: null
-gradient_accumulation_steps: 8
+gradient_accumulation_steps: 1
 eval_accumulation_steps: null
 eval_delay: 0
 learning_rate: 1.0e-04
diff --git a/lazyllm/tools/train_service/client.py b/lazyllm/tools/train_service/client.py
index 9eda8b2e..c78db833 100644
--- a/lazyllm/tools/train_service/client.py
+++ b/lazyllm/tools/train_service/client.py
@@ -79,7 +79,7 @@ def train(self, train_config, token):
                 'num_train_epochs': train_config['num_epochs'],
                 'learning_rate': train_config['learning_rate'],
                 'lr_scheduler_type': train_config['lr_scheduler_type'],
-                'per_device_train_batch_size': train_config['batch_size'],
+                'per_device_train_batch_size': train_config['batch_size'] // train_config['num_gpus'],
                 'cutoff_len': train_config['cutoff_len'],
                 'lora_r': train_config['lora_r'],
                 'lora_alpha': train_config['lora_alpha'],